The First Month with My Hermes AI Agent

Last month I wrote a comparison of personal AI agents from the outside looking in. This post is the other half of that story: what actually happened after I installed one and started handing it real work.

About a month ago I set up Hermes Agent on a Mac mini in my office, connected it to Telegram, and gave it a name — Wrigley. The idea was simple: instead of opening a chat window when I need AI, I wanted an operator that’s always on, remembers everything we’ve done, and can actually execute — run terminal commands, drive a browser, call APIs, schedule its own jobs.

A month in, it has become part of how I run my businesses. Here’s an honest rundown of what we’ve set up and shipped together — for me, and for clients.

The setup

The core arrangement is worth describing because it’s what makes everything else possible:

• Always-on hardware. Hermes runs as a background service on a Mac mini. It survives reboots, restarts itself if it crashes, and is reachable whether I’m at my desk or not.
• Telegram as the front door. My agent lives in my pocket. I send it a message the same way I’d text a colleague, and it answers with full access to its tools — files, terminal, browser, APIs.
• Persistent memory and skills. This is the part that separates an agent from a chatbot. Every workflow we figure out together gets saved as a “skill” — a reusable procedure with the exact steps, credentials locations, and pitfalls documented. The second time we do something, it’s faster. The tenth time, I barely have to explain anything.
• Scheduled jobs. Hermes has its own cron system, so recurring work happens without me asking.

What it does for me every day

Morning briefs, before I’ve had coffee

Two scheduled reports land in my Telegram every morning. At 7am: a stock market brief — futures, movers, anything relevant to positions I follow. At 8am: a business brief that pulls revenue numbers from my Shopify app, ad-network earnings from my content site, advertising spend and profit, and even what the agent itself cost me in API tokens the previous day. One glance and I know where everything stands.

Building that second brief was a real project — it authenticates against multiple dashboards, reconciles ad spend against revenue, and computes actual profit rather than vanity numbers. It’s the kind of report I always meant to build a spreadsheet for and never did.

Content operations on autopilot (with a human veto)

I run a content site in the retirement niche. Hermes now handles a twice-weekly publishing cadence: one slot for rewriting and upgrading older articles, one for drafting brand-new pieces. Every article goes into WordPress as a draft with a preview link sent to me — nothing publishes without my explicit approval. It also handles the unglamorous parts: featured images sized and compressed to the site’s exact spec, SEO-sensible structure, and internal consistency.

That last sentence contains the most important lesson of the month, so I’ll say it plainly: the agent drafts, the human approves. Every workflow we’ve built has that gate.

Ad account research

For the same site, Hermes connects to the Google Ads API to pull search-term reports, analyze keyword performance, and propose new ad groups. It prepares the analysis; I make the changes I agree with. It has also built me profit reporting for a paid campaign on another business, reconciling platform ad spend against actual revenue daily.

Watching the machines

Once you have automation, you need automation that watches the automation. Hermes runs a watchdog over its own scheduled jobs and infrastructure — if a job fails or a service goes down, I get an alert instead of discovering it a week later. It also runs ad-hoc monitors when I ask, like watching a ticket marketplace and pinging me with section and price details when seats matching my criteria appear.

What it has done for clients

This is where it got interesting, because the agent turned out to be a genuine production tool for client work — not just a personal assistant.

This website, actually

The site you’re reading was substantially rebuilt with Hermes doing the heavy lifting. Working over the WordPress REST API — no page-builder clicking — we shipped a new homepage with a full-width hero and lead form, five dedicated service pages with consistent custom headers and mobile navigation, an About page, restyled blog templates, and a Fluent Forms lead pipeline that emails inquiries straight to me. The workflow was the same every time: the agent builds locally, I review screenshots on Telegram, we iterate, and only then does it deploy — draft first when the page is new.

A detail I appreciated: it verifies its own work. After every deploy it checks the live page’s rendered output, computed styles, and mobile behavior against the spec before calling anything done.

A publishing pipeline for a local radio station

One client, a local radio station, needed obituaries posted to their website quickly and accurately from submissions that arrive by email. Hermes now runs a dedicated pipeline: it watches an inbox, drafts each obituary post in the station’s format, sends a preview for approval, and publishes only on a human yes. The same client needed an annual “birthday book” — hundreds of listener birthdays laid out into print-ready sheets — which the agent generated programmatically in a fraction of the time it used to take.

Production support for an Etsy shop

My wife runs an Etsy sticker shop, and Hermes has become part of her production line: batch-processing sticker artwork to the die-cut spec her printer needs, generating consistent product mockups at the right dimensions, and helping with shop presentation like banner design. Repetitive image work that used to eat evenings now takes minutes.

What I’ve learned in month one

• The compounding effect is real. The first WordPress deploy took a whole evening of figuring out API quirks. Now it’s routine, because every lesson got written into a skill. The agent you have in week four is meaningfully more capable than the one you installed in week one — with the same underlying model.
• Guardrails aren’t optional. Everything consequential — publishing, emails, anything customer-facing — requires my approval. That single rule is why I’ve been comfortable expanding its responsibilities. I wrote more about permission design in the previous post, and a month of real use has only reinforced it.
• The value is in the boring work. The flashy demos are code generation; the actual ROI has been reports, image processing, content pipelines, and monitoring — the recurring work that quietly consumes hours.
• It maintains itself, mostly. Updates, health checks, and config migrations are largely self-service — I approve, it executes and verifies. But it’s still software on a real machine, and someone has to be the adult in the room.

Would I recommend it?

If you’re technically comfortable and you have recurring digital work — reporting, content operations, e-commerce back office, client website maintenance — a personal AI agent is worth serious consideration right now. Start read-only, add approval gates for anything that leaves the machine, and let it earn responsibility the way you would with a new hire.

If that sounds valuable but you don’t want to build and babysit it yourself, that’s precisely the kind of automation work we do for clients — from WordPress pipelines to reporting to AI-assisted operations. Tell us about your project in the form above — or get in touch — and I’ll give you an honest read on whether an agent fits your business.